Textual overlap rather than domain alignment: A comparative study of fine-tuning strategies for specialised machine translation with large language models

Lixue Yang, Jiaxin Zhu, Zengxin Zhang · 2026

Read Textual overlap rather than domain alignment: A comparative study of fine-tuning strategies for specialised machine translation with large language models by Lixue Yang, Jiaxin Zhu, Zengxin Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0352256.txt.

Description

Lixue Yang, Jiaxin Zhu, and Zengxin Zhang compare two ways of fine-tuning a large language model for specialist translation, in this 2026 PLOS ONE study. Their test bed was Chinese-to-English political text, an area where general models often stumble. One method retrained every parameter of the model; the other adjusted only a small part, using far less computing power.

On genuinely new political texts, neither method beat the untuned model in any consistent way. Only on test items drawn from the training material itself did both shine — a sign they had memorised overlap, not learned the domain. The sobering lesson: fine-tuning gains can be an illusion created by leaked test data.

Catalog record

CREATORLixue Yang, Jiaxin Zhu, Zengxin Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59457
LICENSECC_BY
RIGHTSCC BY
SOURCEPLOS

Source: https://doi.org/10.1371/journal.pone.0352256

Browse the full library · About the Tower of Babel